Effect of Depth Band Replacement on Red, Green and Blue Image for Deep Learning Weed Detection

Automated agricultural robots are becoming more common with the decreased cost of sensor devices and increased computational capabilities of single-board computers. Weeding is one of the mundane and repetitive tasks that robots could be used to perform. The detection of weeds in crops is now common, and commercial solutions are entering the market rapidly. However, less work is carried out on combatting weeds in pastures. Weeds decrease the grazing yield of pastures and spread over time. Mowing the remaining weeds after grazing is not guaranteed to remove entrenched weeds. Periodic but selective cutting of weeds can be a solution to this problem. However, many weeds share similar textures and structures with grazing plants, making their detection difficult using the classic RGB (Red, Green, Blue) approach. Pixel depth estimation is considered a viable source of data for weed detection. However, systems utilizing RGBD (RGB plus Depth) are computationally expensive, making them nonviable for small, lightweight robots. Substituting one of the RGB bands with depth data could be a solution to this problem. In this study, we examined the effect of band substitution on the performance of lightweight YOLOv8 models using precision, recall and mAP50 metrics. Overall, the RDB band combination proved to be the best option for YOLOv8 small and medium detection models, with 0.621 and 0.634 mAP50 (for a mean average precision at 50% intersection over union) scores, respectively. In both instances, the classic RGB approach yielded lower accuracies of 0.574 and 0.613.